wisconsin.txt

After more than 10 years in administration, including 9 as Dean of
Arts and Sciences and 1 as interim Provost at UConn, I have returned
to my faculty position.  I am spending a year as a visiting scientist
at the Jackson Laboratory for Genomic Medicine (JAX-GM) in Farmington,
Connecticut, trying to get a grip on some of the mathematical problems
of interest to researchers in cancer genomics.  In this talk, I will offer some personal
observations about being a mathematician and a high-level administrator, talk a bit about
the research environment at an independent research institute like JAX-GM, outline
a few problems that I've begun to learn about, and conclude with a
discussion of how these experiences have shaped my view of graduate training in mathematics.

Observations as an administrator
Challenges/pleasures of administration:
		     intellectual diversity
		     other types of diversity
		     truly intractable problems
		     teamwork
		     opportunity to be a force for good
issues:
	- doubts about  the curriculum
	- isolation of the department
	- stubbornly slow process of diversification
	
broader observations:
	-- length of academic career is a challenge for many people; change can be good.
	-- relative impermeability of academic mathematics


Jackson Labs

Founded 1929 in Bar Harbor Maine
Originally supported by Edsel Ford and Roscoe Jackson (Hudson Motor cars)
Home of mouse genetics
26 Nobel Prizes associated with JAX
George Snell, JAX Faculty, won in 1980 for genetics of the immunte system.
Huge business in mice  -- 3M mice per year
88M per year grant support
256M per year in the mouse business and other services
Mouse genome database
New facility in Connecticut dedicated to 'precision medicine'
 -- 350 employees
 ~ 30 faculty
 huge technical support infrastructure
  -- information technology and "computational services"
  -- imaging
  -- sequencing
  
Math departments think of NSA or Finance as places where there are jobs for PhD's.
Place like JAX offers a research environment (no teaching) but can be a steady job -- compare
for example with a permanenent teaching faculty job.

Here's a sample job ad:

Responsibilities

The successful candidate will develop and apply computational methods
for genomic data analyses, have novel opportunities at the
intersection of algorithms, data interpretation, and experimental
design in close collaboration with the experimental team in the
lab. The applicant is expected to have Ph.D. in computational biology,
cancer genomics, or a similar discipline, and demonstrate strong
computing expertise, publication, and communication skills. We also
consider exceptional applicants who are new to computational biology,
including statistics, applied math, physics, and computer science.

Qualifications

    PhD. in a quantitative area: computational biology, computer science, statistics, physics and applied math

    Programming in R, Python, Perl or C++

    Experience in genomic dataset analysis and integration

    Three (3) years of postdoctoral training or equivalent research experience, and demonstrated meritorious scientific performance.

    Strong publication record (two or more peer-reviewed publications).

    Ability to integrate and apply feedback in a professional manner, prioritize and drive to results with a high emphasis on quality, and work as part of a team.

    Proven ability to analyze and synthesize information, and demonstrated resourcefulness in finding appropriate solutions to problems and initiative in presenting alternatives and implementing	solutions to ensure effective change and/or eliminate or mitigate potential negative effects.

    Exceptional organizational and project management skills, including the ability to plan, schedule, and successfully carry out multiple projects at the same time to successful conclusion with	 minimal supervision.

Quality orientated including delivering accuracy and attention to detail.


Some examples of people at JAX -- many non-biologists -- changing field draws in people from lots of areas.

Jeff Chuang (Assoc Prof, faculty member) PhD from MIT in statistical physics
Bill Flynn, "Application computational scientist", PhD in Physics (large molecules) from Rutgers
Vida Ravanmehr, PhD in Electrical/Computer Engineering (working on error correcting codes) BA/MA in Applied Math
Peter Robinson (Professor), BA and MS in Computer Science, MD from Penn.

But not many math PhD's (pure or applied) overall.

What are they working on?

Many different things, but here are two examples of problems.

1. (Robinson group)  The Human Phenotype Ontology (HPO)

Phenotype = symptom.  HPO is a large directed acyclic graph, where each node represents a symptom/phenotype and they are classified
from general to specific.  Items ("terms") in the HPO are curated by experts.
For example:
[Term]
id: HP:0010082
name: Symphalangism affecting the distal phalanx of the hallux
synonym: "Fused outermost bone of big toe" EXACT layperson [ORCID:0000-0001-5208-3432]
xref: UMLS:C4024063
is_a: HP:0001859 ! Distal foot symphalangism
is_a: HP:0010053 ! Abnormality of the distal phalanx of the hallux
is_a: HP:0010091 ! Symphalangism affecting the proximal phalanx of the hallux
is_a: HP:0010191 ! Symphalangism affecting the distal phalanges of the toes
created_by: doelkens
creation_date: 2009-05-29T12:16:28Z

DAG is annotated by (mendelian genetic) diseases (So a disease is a subtree of the ontology)

Possible applications:
useful for clinical studies because it nails down symptoms in a systematic nomenclature.
Can be used for more interesting problems --  given a set of symptoms, find potential diseases consistent with those symptoms.
This type of reasoning problem from ontologies has a long history.
(give examples)

Also used to identify cross-species similarities and to find appropriate animal models
And to try to pair specific genetic variants with diseases.

2.  (Chuang group) broader problem: tumor evolution.  Typical tumor is made up of subpopulations of cells with common
genetic profile.  Some of these subpopulations might be susceptible to treatment with a particular chemotherapy agent
to which others are resistant.

Patient-derived xenografts (PDX): take tumors from people, grow them in immuno suppressed mice so there is space to experiment on them.

Some attempts to reconstruct history of cell populations using ideas from evolutionary biology, but still an active area of research

Management of PDX studies is a large complicated business.

Key tool is sequencing -- different types -- DNA, RNA, single-cell.
My current project is to look at issues in single cell RNA sequencing, so here is a brief overview.

Recall how genetics works: DNA -> mRNA -> protein
Different cells in your body have same* DNA but different expression profiles -- different levels of mRNA and so different collection of proteins.

Prior technology (RNA-seq) looked at a sample from a mixture of cells and computed the relative abundance of different mRNA molecules.
So a typical RNA-seq experiment might take 3 replicates of a control and 3 of a treatment (say drugs or whatever) and compute the expression levels
of the genes in the two cases, then try to draw conclusions.  So you have a matrix with 30000 genes and six samples.  Various strategies to extract
meaning from this situation.

single-cell works differently.  Cells get sequenced one at a time and you get an expression profile for all of the genes for each cell.
Result might be an integer matrix of 3000 cells by 30000 genes.  Each integer counts the number of mRNA associated with that gene.

BUT: the data is confounded because the small amounts of material in each cell mean that many things DON't get counted that should.  So the
matrix has lots and lots of zeros.

A typical goal is:
  -- to cluster the cells by their expression profiles;
  -- identify the genes whose expression levels characterize the clusters;
  -- find a biological interpretation for these genes.

In the context of tumor heterogeneity, perhaps the clusters might be associated with certain biochemical pathways that can be targeted with particular drugs.

Example of one  associated statistical model (there are many proposals for this):

Typical approach would be to use principal componenet analysis or some fancier method to reduce the dimension and identify clusters.

Problem: How to deal with the issues of:
	 -- normalization of the data 
	 -- what effect does the dropout have on the dimensionality reduction?
	 -- should one try to impute the zero values?  if so, how?